Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office Action has been withdrawn pursuant to 37 CFR 1.114.
Detailed Action
This Non-Final Office Action is responsive to Applicants’ RCE submission dated 5/7/26, including amendments and arguments previously received 4/27/26. Claims 1-8, 10-17, and 19-22 remain pending, of which claims 1 and 16-17 are independent.
Claim Rejections - 35 USC § 101
4. 35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
5. Claims 1-8, 10-11, and 14-16 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Taking independent claim 1 as representative, Step 1 of the Subject Matter Eligibility Test (MPEP 2106) determines whether the claimed invention is directed to a process, machine, manufacture or composition of matter. In this instance, the claim is directed to “...” which qualifies as “...” under the Step.
Moving on to Step 2A Prong One, we determine whether the claim recites an abstract idea, law of nature, or natural phenomenon. MPEP 2106.04 (I-II) and 2106.04(a).
Claim 1 recites, in part, the following as reproduced just below. The Examiner has bolded the limiting features that are directed to an abstract idea (to be discussed in 2A Prong One) and underlined the features that are directed to an additional element (to be discussed in 2A Prong Two and 2B):
A method implemented via a computing device, the method comprising:
receiving, by the computing device, data associated with a subject having a neurodevelopmental disorder (NDD), the data associated with the subject comprising demographic data, schooling data, family medical data, prior therapy data, observational assessment data, medication data, goals data, or combinations thereof, wherein a neurodevelopmental disorder treatment recommendation (NDDTR) model is associated with the computing device; and
evaluating, by the NDDTR model associated with the computing device, the data associated with the subject, wherein the NDDTR model is configured to evaluate the data associated with the subject to determine a therapy recommendation, wherein evaluating the data associated with the subject comprises transforming the data into discrete numerical vectors, wherein the discrete numerical vectors are provided to the NDDTR model to determine the therapy recommendation, wherein transforming the data into discrete numerical vectors comprises converting categorical and text data associated with the subject into binary inputs and encoding the binary inputs as numerical vectors, wherein the NDDTR model is a machine learning model selected from the group consisting of a deep learning model, a generative adversarial network, a computational neural network model, a recurrent neural network model, a perceptron model, a classical tree-based machine learning model, a decision tree type model, a regression type model, a classification model, a reinforcement learning model, and combinations thereof, wherein the NDDTR model is configured to output a score between 0 and 1, wherein the score is compared to an operating threshold to determine the therapy recommendation, wherein the operating threshold is effective to yield a sensitivity of from 0.71 to 0.99, wherein the NDDTR model is configured to prioritize true positives and limit false negatives, and wherein the therapy recommendation comprises a standard of care associated with provision of applied behavior analysis (ABA) therapy.
In view of the bolded features noted above by the Examiner, the claim is principally directed to a method that involves, having received a specified type of data, using a model to evaluate that data to determine a recommendation. Characterized this way, the claim’s bones are largely that of an abstract idea, that is further clarified via additional elements, with the abstract idea essentially being an evaluation or judgment or mental steps that can be performed by the human mind. The claim does clarify that the model’s evaluation involves transforming the data into numerical vectors, which further comprises transforming categorical/text data into binary inputs to be encoded in that same vector-based way. This too is achievable using the human mind and/or pen and paper, since it is essentially transforming data from one format to another. Alternatively, these same features can be understood to be mathematical operations, which are also understood to be an example of an abstract idea. Under this rationale, the bolded features noted above direct the claim’s interpretation to be that of largely an abstract idea, with some additional elements (as underlined) included which will be discussed in the analysis to come.
Moving on to Step 2A Prong Two, we determine whether the claim recites additional elements that integrate the judicial exception into a practical application. MPEP 2106.04(d). The additional elements are as noted below:
The use of a computing device and further the association between that computing device and the model is understood to implement what is essentially the abstract idea as discussed above via a general purpose computer. Said another way, these limitations are akin to merely reciting the words “apply it” or its equivalent in relation to the abstract idea as disposed into a generalized computing environment, i.e., merely using a computer as a tool to perform an abstract idea. Hence, this is not sufficient to integrate the judicial exception into a practical application.
The Examiner notes that, by way of the amendment, the model has been clarified to be a machine learning model selected from a list of different machine learning model types. While this does provide some additional detail, the breadth of the recitation suggests that the type of model is not meaningfully important since pretty much every imaginable type of machine learning model is included, and hence it again begs the question as to whether the claim simply is taking the limitation or step of an evaluation or judgement or mental step and saying “apply it” to a generalized computer environment. Hence, because it is understood as merely applying an abstract idea to a computer-implemented environment, and is therefore not sufficient to integrate it into a practical application.
The data associated with the subject, e.g. as the model is applied to, is further clarified to be one or many types of categorical data, e.g. demographic, school, family medical, etc. While this does provide some additional detail, simply clarifying what data in terms of type or kind is being subject to the abstract idea, evaluation, judgment, mental process, etc. does not make the abstract idea any less abstract and does not meaningfully integrate it into a practical application. At the end of the day, the critical feature is still an evaluation, judgment, or the like being applied to a particular type of data, and hence the principal thrust of the claimed invention is still largely abstract.
The step for receiving the data is understood to be an example data collection, gathering, etc. and hence insignificant extra-solution activity that is incidental to the abstract idea itself, and therefore not sufficient to integrate the abstract idea into a practical application.
Additional elements clarify the model’s output range as being a score between 0 and 1, a model’s sensitivity setting/calibration relating to the score/output, and a prioritization of the model for true positives over false negatives. Calibrating a model’s output/score to a certain range is essentially a mathematical adjustment/transformation or similar detail, and can otherwise be understood to be a simple mental step. Tuning a model’s sensitivity and prioritizing its calibration to favor true positives over false negatives appear to be details that relate to how the model is constructed, whether it’s the steps included within the model or how the inputs or other parameters are biased/adjusted. These are evaluations or judgments that go into how the model is built, for example with a sort of evergreen aim relating to improving a model’s performance or accuracy. Hence, these additional elements cannot be understood to overcome the general characterization of the claimed invention being largely directed to an abstract idea and one not integrated into a practical application.
Finally, in Step 2B, we evaluate the additional elements to determine whether they amount to significantly more than the judicial exception. MPEP 2106.05. Revisiting the additional elements as discussed above per Step 2A Prong Two, the Examiner does not believe the additional elements are persuasively sufficient to provide significantly more than the abstract idea as otherwise characterized by the Examiner. The Examiner reiterates the reasoning relating to the secondary elements as provided above in the prior step:
A generalized use of a computer with respect to a model’s application to subject data appears to be nothing more than applying the model to a computer, and cannot be understood to provide significantly more than the abstract idea.
The receiving of the data as insignificant extra-solution activity cannot be understood to provide significantly more than the abstract idea as applying a model to that same data.
The type of the model and its calibration when relating to sensitivity considerations are meaningful to how the model performs and its accuracy, for example, but if the model itself is largely an abstract idea in terms of providing evaluation, judgement, and mental steps type processing, then these features that further define the model are not meaningfully persuasive to provide something significantly more than the model and those abstract idea aspects therefor.
Based on the reasoning as provided above, independent claim 1 is not eligible subject matter.
Independent claim 16 includes many of the same or similar limitations as claim 1, and are therefore rejected under the same rationale.
Further, claim 16 is expressly directed to a computing system, which the Examiner understands to be a type of “machine” under Step 1. The additional context provided by the system per claim 7 (e.g., processor, non-transitory computer-readable medium) does not meaningfully alter the result of the Examiner’s analysis as provided above per claim 1; for example, the computer-relatedness of the claim as recited at this level of generality is not “a particular machine that is integral to the claim” per MPEP 2016.04(d)(2). Rather, merely reciting the words “apply it” or their equivalent with a judicial exception, or merely including instructions to implement an abstract idea on a computer or machine, or merely using a computer or a machine as a tool to perform an abstract idea, has been found not sufficient. See, e.g., MPEP 2106.05(f).
The other dependent claims which depend from independent claim 1 are likewise rejected, because they are not deemed to materially change the outcome of the analysis as provided for the independent claims. The Examiner will address each dependent claim below, in turn:
Regarding claim 2, the claim is directed to a clarification as to what type of subject is subject to analysis via data gathering/collection and evaluation thereof. At best, the limitation clarifies a descriptive scope for the gathered/collected data without any sort of meaningful further active limitation that would serve to integrate the abstract idea of claims into a practical application or otherwise provide significantly more to the abstract idea. Merely clarifying the type of data giving to an evaluation, judgment, or mental process does not itself make that evaluation, judgement, or mental process any more concrete in a way that overcomes the subject matter eligibility rejection as presented.
Regarding claims 3-7, the claim is directed to a clarification as to define what type of data is subject to gathering/collection and evaluation thereof. At best, the limitation clarifies the type of data used as inputs to serve as a basis for realizing an output result of a mental process such as the one discussed above per claim, and hence does not constitute any sort of meaningful further active limitation that would serve to integrate the abstract idea of claims into a practical application or otherwise provide significantly more to the abstract idea.
Regarding claim 8, the claim is directed to a clarification as to define the richness or scope of what data is subject to gathering/collection and evaluation thereof. At best, the limitation clarifies aspects of the data (e.g., metadata essentially) used as inputs to serve as a basis for realizing an output result of a mental process such as the one discussed above per claim, and hence does not constitute any sort of meaningful further active limitation that would serve to integrate the abstract idea of claims into a practical application or otherwise provide significantly more to the abstract idea.
Regarding claim 10, the claim is directed to a clarification as to what type of model is used to implement the evaluation and recommendation features of the claimed invention. The Examiner reasons that the different types of models as recited provide differentiation as to what machine learning approach is used in a very general and high level of detail, which amounts to essentially what algorithms or steps are used to translate a set of inputs to a set of outputs. While this is a meaningful distinction in defining the model of the claimed invention, the limitation does not constitute any sort of further steps that would actively integrate the abstract idea claims into a practical application or otherwise provide significantly more to the abstract idea. At best, it merely serves to define what algorithm or what mental process constitutes the model and hence the claimed invention.
Regarding claim 11, the claim is directed to a particular type of machine learning model/approach, and hence the rationale provided above per claims 9-10 are reiterated here.
Regarding claim 14, the claim is directed to clarifying a definition for what the output is based on the model’s evaluation of the inputs. While this is a meaningful distinction in defining the model of the claimed invention, the limitation does not constitute any sort of further steps that would actively integrate the abstract idea claims into a practical application or otherwise provide significantly more to the abstract idea. At best, it defines a context or a type of information that is the output of an evaluation or judgement.
Regarding claim 15, the claim is directed to a highly-generalized step of acting upon the model’s output by providing therapy pursuant to the recommendation. Without further detail, merely following through on an evaluation, judgment, or mental process result, as is recited here, constitutes extra-solution activity that is incident to the abstract idea itself, and does not meaningfully integrate it into a practical application or provide significantly more at this present level of detail.
6. A comment from the Examiner: independent claim 1’s model appears to be already learned/trained, as relating to the active scope of the claim. Said another way, the claim involves applying data to a model that is otherwise static or fixed, and for that reason, the model is more readily equated to a fixed evaluation, judgment, or mental process framework, even with the detail provided by Applicants in the rest of the claim and in the dependent claims. The Examiner’s consideration of this claim would change if the claimed subject matter could be understood to be improving the operation of the computer/machine or the technological field, and this is sometimes achievable by presenting details within the claim as to how the model’s building, training, and so forth is an actual technical improvement that improves the model’s performance relative to itself, the state of the art, etc. While the Examiner appreciates the Applicants’ comments provided on pages 13-14 of the reply received 4/27/26, they do not appear to persuasively make this case in a way that the Examiner believes would overcome the 101 rejection.
Applicants’ aforementioned recent arguments, as found on pages 13-14 of the reply received 4/27/26, appear to essentially argue that (i) the claimed invention requires too many steps to be reasonably abstract because it cannot be reduced to an evaluation, judgment, or mental process performable in the mind of a person, and also that (ii) the claimed invention itself as being facilitated by a trained model is then a concrete technical improvement over a traditional clinical application of such judgments, evaluations, and the like. To these arguments, the Examiner respectfully disagrees:
One, the Applicant provides no detail as to how massive the data involved is, and for all the Examiner can interpret, the data may be manageable to be worked out by a person mentally and/or via pen and paper. For example, Applicants’ claim itself includes the possibility that the model is “a decision tree type model” or even a “classical tree-based machine learning model”, which the Examiner interprets to be one or more decision trees or having a tree structure, e.g. flowcharts or an ensemble thereof. If the model can simply be just that in satisfaction of the claim’s requirements, then the argument that the model is not implementable as a tree or flowchart as might be implemented with pen and paper or on a whiteboard in a lab or office is not very persuasive.
Two, the recitation of the model itself per the independent claims is at a high level, both in terms of level of detail and breadth, such that it invites the interpretation that the details of the model itself do not matter significantly; in which case, again, is this not just an abstract idea being subject to an “apply it” rationale in relation to a general purpose computer? Applicants appear to argue that the claimed invention is specific in this regard, and hence argue that the rejection under 101 is inappropriate, but yet provide a high level and extensive laundry list of technical models that would suffice to implement the invention.
Claim Rejections - 35 USC § 103
7. In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
8. The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office Action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
9. The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
10. Claims 1-3, 5-8, 10, and 14-16 are rejected under 35 U.S.C. 103 as being unpatentable over U.S. Patent Application Publication No. 2023/0045696 (“Griffin”) in view of Non-Patent Literature “Machine learning-based ABA treatment recommendation and personalization for autism spectrum disorder: an exploratory study” (“Kohli”) and further in view of Non-Patent Literature “Model Selection” (“Datasklr”), Non-Patent Literature “3.3 Metrics and scoring: quantifying the quality of predictions” (“Scikit”), and Non-Patent Literature “3 Ways to Encode Categorical Variables for Deep Learning” (“Brownlee 1”).
Regarding claim 1, GRIFFIN teaches a method implemented via a computing device (a computing device as taught per FIG. 1 and [0023]-[0029] (“health evaluator” that includes memory, control circuitry, a UI, and so on, which may be implemented using a tablet or a smartphone as discussed per [0025])), the method comprising:
receiving, by the computing device, data associated with a subject having … a healthcare condition (FIG. 2 step 22, as discussed per [0068] discussing the generation and receipt of health data as provided to the health evaluator, where the health data is used for evaluation, prediction, and recommendation aspects relating to a patient’s health ([0066])), the data associated with the subject comprising demographic data ([0033]: “The pertinent health data can include patient data, which is data about the specific patient (e.g., age, weight, height, ethnicity, race, economic status, lifestyle factors ...” and later “professional, marital status, age, sex, race, household factors ...”, where the Examiner equates the italicized portions discussed here per [0033] with the recitation for “demographic data”), schooling data, family medical data, prior therapy data ([0033]: “treatment data, which is data regarding treatment actions for the patient (e.g., one or more of lifestyle changes, additional or different medications, one or more procedures, frequency of follow ups, etc.)” as used for evaluation, prediction, and recommendation aspects relating to a patient’s health), observational assessment data, medication data ([0020]: “prior medications” are used to analyze the patient’s state of health), goals data, or combinations thereof, wherein a ... treatment recommendation (TR) model is associated with the computing device (the “health evaluator” as taught, e.g. in relation to FIG. 2 as discussed in earlier mappings provided just above, is situated in the taught framework relative to a computing device as taught per FIG. 1 and [0023]-[0029]); and
evaluating, by the … treatment recommendation (TR) model associated with the computing device, the data associated with the subject, wherein the TR model is configured to evaluate the data associated with the subject to determine a therapy recommendation ... wherein the therapy recommendation comprises a standard of care (FIG. 2 steps 24-30 culminating with an analysis result that is “treatment information”, which [0085] clarifies to include “... health actions that correspond with improved patient health and/or with the patient achieving the modified health data associated with improved patient health. For example, the health evaluator can output one or more of treatment actions (e.g., prescriptions, doctor appointments, etc.) and patient actions (e.g., at least 30 minutes of exercise, diet changes, etc.)”, where the Examiner reasons that the taught “treatment actions” and “patient actions” read on the recited “therapy recommendation” comprising “a standard of care”) associated with provision of applied behavior analysis (ABA) therapy ([0110]: “The health evaluator 1 and method 32 provide significant advantages. The health evaluator can predict lab results and generate an accurate medical diagnosis for a patient based on the current health of the patient and based on modifications in the health of the patient. The modifications can provide information regarding treatment actions, patient actions, etc. that will result in an optimal patient health outcome. By using the predictive model to predict how patient behavior, medications and interventions affect future vital signs and lab results, in order to predict future diagnoses, the most beneficial patient behavioral changes, medications and interventions can be identified that will also minimize the number of diagnoses of that patient. The predictive health evaluator model is configured to predict how patient behavior, medications, and interventions affect future vital signs and lab results, in order to predict future diagnoses, the most beneficial patient behavioral changes, medications, and interventions that provide an optimal treatment plan. The health evaluator analyzes all pertinent health data to generate predictive results, providing decision support to clinicians that may not be fully aware of other actions regarding the patient's health. Individual clinicians can analyze and modify suggested treatments based on the outputs from health evaluator thereby accounting for all aspects of the patient's health.”) and
wherein the … TR model is a machine learning model selected from the group consisting of a deep learning model, a generative adversarial network model, a computational neural network model, a recurrent neural network model, a perceptron model, a classical tree-based machine learning model, a decision tree type model ([0050] discussing that the health predictor model can be an ensemble model of decision trees, as part of a larger discussion of the tree based model per [0050]-[0060]), a regression type model, a classification model, a reinforcement learning model, and combinations thereof ...
Griffin does not teach the further limitations of a subject having a healthcare condition that is a neurodevelopmental disorder (NDD) thereby making the treatment recommendation model per Griffin specifically a neurodevelopmental disorder treatment recommendation (NDDTR) model. Griffin notes a scope for its recommendation model that encompasses an open-ended breadth for diseases/conditions ([0117]) for which there could be patient health data subject to analysis thereof, but the open-ended list does not explicitly include NDDs as is recited. That said, the Examiner believes that NDDs, like many other conditions such as those per Griffin’s [0117], can be subject to characterization and model-driven analysis for recommendation purposes as Griffin does more broadly. The Examiner relies upon KOHLI to teach what Griffin explicitly lacks, see e.g. Kohli’s comparable recommendation system that considers patient similarity (section 2.3), using patient data, to generate a personalized treatment recommendation (sections 2.4-2.5) for a particular patient. See also its FIG. 1 on page 6 for a system/flow diagram. See Abstract, and also item #2 of participant inclusion criteria (found at the top of page 5): “Children should have a diagnosis of autism spectrum disorder using standardized instruments such as the DSM-V, CARS-2, ADI-R, INDT-ASD, ISAA, or any other evidence-based ASD diagnostic tool.”
Both Griffin and Kohli involve data-driven and model-based approaches to providing a personalized treatment recommendation to health care patients. Hence, they are similarly directed and therefore analogous. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to extend Griffin’s open-ended framework to include autism as a considered subject matter, with a reasonable expectation of success, e.g., to address the shortage of qualified personnel to help patients with an ASD diagnosis (as Kohli addresses in its Introduction section) using Griffin’s comparatively more deployment-ready and treatment-forward framework (e.g., relative to Kohli’s posture as more of a study).
The claimed invention has been further clarified to recite the following additional limitations:
“wherein evaluating the data associated with the subject comprises transforming the data into discrete numerical vectors, wherein the discrete numerical vectors are provided to the NDDTR model to determine the therapy recommendation” and relatedly “wherein transforming the data into discrete numerical vectors comprises converting categorical and text data associated with the subject into binary inputs and encoding the binary inputs as numerical vectors.”
“wherein the NDDTR model is configured to output a score between 0 and 1, wherein the score is compared to an operating threshold to determine the therapy recommendation, wherein the operating threshold is effective to yield a sensitivity from 0.71 to 0.99, wherein the NDDTR model is configured to prioritize true positives and limit false negatives.”
Regarding these additional limitations, the Examiner believes they are broadly known in the state of the art pertaining to machine learning:
Specifically regarding the limitations grouped into the 1st bullet point as provided just above, BROWNLEE 1 teaches the encoding of categorical information for deep learning, where categories may be associated with an integer, and/or is mapped to a binary vector, and so forth. See the reference’s page 2, under the heading The Challenge with Categorical Data. Hence, the Examiner reasons that this teaching can be readily applied to encode, in the manner Applicants recite, at least some of the “data associated with the subject [that is] demographic data”, which the Examiner had relied upon Griffin’s [0033] to teach: “The pertinent health data can include patient data, which is data about the specific patient (e.g., age, weight, height, ethnicity, race, economic status, lifestyle factors ...” and later “professional, marital status, age, sex, race, household factors ...” Some of this demographic data per Griffin is clearly categorical (e.g., ethnicity, race, martial status, sex, race, etc.), and hence could reasonably be provided by a subject in a text-based input rather than a numerical-based way, and as such is amenable to Brownlee 1’s encoding approach to prepare it for processing by a machine-learning framework such as Griffin’s or especially Kohli’s.
Regarding the limitations grouped into the 2nd bullet point as provided just above, SCIKIT teaches generally, in its section 3.3, the differing approaches to scoring the output of a predictive model, e.g. like Griffin’s or Kohli’s the Examiner reasons. More specifically, section 3.3.2.4 teaches a particular scoring approach, such that the score ranges from 0 to 1, that balances accuracy by considering true positive and true negative rates. The Examiner notes that the balanced accuracy equation considers true positive and false negative cases. Intuitively, any classifier or predictive model that is striving for accuracy, e.g. the objective of basically any such modelling, would prioritize true positives and would deprioritize false negatives as the former promotes accuracy and the latter is damaging to accuracy. Hence, the Examiner reasons that a model, such as Griffin’s or Kohli’s, which is tuned/optimized for accuracy using a scoring approach as taught here by Scikit, would emphasize these same objectives in the producing of an improved/preferred model. In that sense, the Examiner reasons Scikit’s teaching here can be understood to be an operating threshold that provides or captures a sensitivity for the model that satisfies the recited limitation’s goals/aims.
However, to the extent that Scikit does not sufficiently/concretely teach sensitivity relating to true positive and false negatives in a numerical manner, the Examiner reasons that DATASKLR’s F1 Score perhaps more succinctly provides such a teaching of a metric that can be evaluated for a model to promote those same goals/aims: “The F1 Score is the weighed average of precision and recall, hence it takes both false positives and false negatives into account”
Brownlee 1, Datasklr, and Scikit all teach numerical approaches to evaluates a model’s performance, in terms of its accuracy and so forth and by taking into account the same or similar characterizations of true/false positives/negatives when it comes to the model’s output verses a ground truth. Hence, they are similarly directed to improving predictive/classifying modelling frameworks, such as those taught by Griffin and Kohli. It would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to apply the teachings of Brownlee 1, Datasklr, and Scikit to a combined framework of Griffin and Kohli with the purpose of evaluating and improving the model’s performance.
Regarding claim 2, Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 teach the method of claim 1, as discussed above. The aforementioned references further teach the additional limitations wherein the neurodevelopmental disorder is autism spectrum disorder (ASD) (Kohli: Abstract, and also item #2 of participant inclusion criteria (found at the top of page 5): “Children should have a diagnosis of autism spectrum disorder using standardized instruments such as the DSM-V, CARS-2, ADI-R, INDT-ASD, ISAA, or any other evidence-based ASD diagnostic tool.”). The motivation for combining the references is as discussed above in relation to claim 1.
Regarding claim 3, Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 teach the method of claim 1, as discussed above. The aforementioned references further teach the additional limitations wherein the data associated with the subject comprise the demographic data, wherein the demographic data comprise age data (Griffin’s [0033]: “The pertinent health data can include patient data, which is data about the specific patient (e.g., age, ...)” and later “professional, marital status, age, sex, race, household factors ...”; and Kohli: page 2, just before the start of section 2 discussing “We used assessment records, sociodemographic data such as age and gender ...”, and page 4, section 2.5, teaching “The patient’s age and vitals are analyzed to identify similar patients.”, and section 3.1, teaching “The sociodemographic data, including age, gender, ethnicity, and ASD diagnosis for participants, were collected”, and page 5, section 3.3 teaching “... We assume that children would benefit from ABA treatment goals that have shown success to an existing child with similar gender, age, and assessment scores ... ”). The motivation for combining the references is as discussed above in relation to claim 1.
Regarding claim 5, Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 teach the method of claim 1, as discussed above. The aforementioned references further teach the additional limitations wherein the data associated with the subject comprise the prior therapy data, wherein the prior therapy data comprise an indication of the subject having previously received occupational therapy, an indication of the subject having previously received speech therapy, an indication of duration of applied behavioral analysis (ABA) therapy previously received by the subject (Kohli: page 7, section 3.6: “Therefore, we investigated the relationship between patients’ treatment profiles, including sociodemographic data (age, gender), domain and target codes, treatment duration, and effectiveness (days to mastery) as interaction items and as an input vector to develop an effective treatment recommendation system using CF.”), an indication of amount of ABA therapy previously received by the subject (Kohli: page 7, section 3.6: “Therefore, we investigated the relationship between patients’ treatment profiles, including sociodemographic data (age, gender), domain and target codes, treatment duration, and effectiveness (days to mastery) as interaction items and as an input vector to develop an effective treatment recommendation system using CF.”), or combinations thereof. The motivation for combining the references is as discussed above in relation to claim 1.
Regarding claim 6, Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 teach the method of claim 1, as discussed above. The aforementioned references further teach the additional limitations wherein the data associated with the subject comprise an indication of the subject's tendency toward aggressive behavior, an indication of the subject's tendency toward stereotypy, an indication of the subject's tendency toward destructive behaviors, an indication of the consequences implemented by a caregiver of the subject responsive to negative behavior, an indication of the subject's ability to be understood, an indication of the subject's ability to understand others (Kohli’s patient similarity is based on scoring using assessments from SRS-2 and VB-MAPP, as discussed on page 7, column 1, paragraph 1, and further down the same page in the table for Algorithm 1 (see similar teaching in section 3.6’s first paragraph and in the last paragraph of section 5.1), and where the SRS-2 and VB-MAPP assessments (per Appendix 1 on pages 21-22) are understood to indicate “social communication” and “verbal and related skills ... measuring learning and language milestones” and “language acquisition”), an indication of variety of foods eaten by the subject, an indication of the subject's ability use a toilet independently, an indication of the subject's ability to bathe independently, an indication of stimulatory behaviors exhibited by the subject, or combinations thereof. The motivation for combining the references is as discussed above in relation to claim 1.
Regarding claim 7, Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 teach the method of claim 1, as discussed above. The aforementioned references further teach the additional limitations wherein the data associated with the subject comprise the goals data (Kohli: page 5, section 3.2 teaching “Parents used mobile or web applications to track their child’s progress, shared 10–15 min child’s progress videos weekly, and recorded responses to skill development treatment goals. At the start of months zero, four, and six, the children underwent a detailed SRS-2 and VB-MAPP assessment.”), wherein the goals data comprise an indication of a goal of improved communication skills (where the SRS-2 and VB-MAPP assessments (as clarified per Appendix 1 on pages 21-22) are understood to indicate “social communication” and “verbal and related skills ... measuring learning and language milestones” and “language acquisition”, i.e. the Examiner reasons that the goals as previously referenced when understood in relation to those skill assessment tools particularly are understood to include communication goals/skills), an indication of a goal of improved diet, an indication of a goal of increased independence, an indication of a goal of improved ability to express emotions, of combinations thereof. The motivation for combining the references is as discussed above in relation to claim 1.
Regarding claim 8, Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 teach the method of claim 1, as discussed above. The aforementioned references further teach the additional limitations wherein the data associated with the subject comprise structured data (electronic medical records, per Griffin’s [0017] and Kohli’s sections 2.3-2.4 and 3.5, e.g. as maintained in a database as shown in Kohli’s FIG. 1 on its page 6, would be understood to constitute “structured data” as recited), wherein the data associated with the subject comprise a plurality of data features, and wherein the plurality of data features comprises not more than 30 different data features (features per Griffin’s EMRs, as discussed at [0039]: “Each set of features can include tens, hundreds, or thousands of features”, i.e., the number of features can constitute a number in an incredibly large range that could be less than about 30 if just 2-3 instances of “tens ... of features” as Griffin teaches). The motivation for combining the references is as discussed above in relation to claim 1.
Regarding claim 10, Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 teach the method of claim 1, as discussed above. The aforementioned references further teach the additional limitations wherein the machine learning model is a gradient-boosted tree model comprising a plurality of weighted decision trees (Griffin’s [0046]: decision trees and boosted decision trees, and [0050]: “gradient boosted decision tree”, and [0056]: “The decision trees are weighted based on the predictive accuracy ...”). The motivation for combining the references is as discussed above in relation to claim 1.
Regarding claim 14, Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 teach the method of claim 1, as discussed above. The aforementioned references further teach the additional limitations wherein the standard of care associated with the provision of the ABA therapy comprises an indication of intensity of ABA therapy, an indication of ABA therapy services (Griffin: FIG. 2 steps 24-30 culminating with an analysis result that is “treatment information”, which [0085] clarifies to include “... health actions that correspond with improved patient health and/or with the patient achieving the modified health data associated with improved patient health. For example, the health evaluator can output one or more of treatment actions (e.g., prescriptions, doctor appointments, etc.) and patient actions (e.g., at least 30 minutes of exercise, diet changes, etc.)”, where the Examiner reasons that a duration of exercise as taught is an indication of intensity as recited, and further where the Examiner reasons that the taught “treatment actions” and “patient actions” read on services as recited, e.g. exercise and diet plans for which services are known to be involved at a user’s discretion, e.g. via a gym, a trainer, a dietician, a nutritionist, a chef, etc. (meaning, it would be obvious to employ a service such as those mentioned just now to pursue the recommend treatment plan)), or an indication of one of a comprehensive ABA therapy or a focused ABA therapy. The motivation for combining the references is as discussed above in relation to claim 1.
Regarding claim 15, Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 teach the method of claim 1, as discussed above. The aforementioned references further teach the additional limitations further comprising providing therapy to the subject based upon the therapy recommendation (a recommendation of “personalized treatment prescriptions” per Kohli’s Introduction on its page 2 but see also Kohli’s sections 2.4-2.5 and 3.2 discussing more extensively what the personalized treatment could entail (which the Examiner reasons sufficiently reads on the high-level “providing therapy” of this instant limitation)). The motivation for combining the references is as discussed above in relation to claim 1.
Regarding claim 16, the claim includes the same or similar limitations as claim 1 discussed above, and is therefore rejected under the same rationale. The instant claim specifically recites a processor and a non-transitory computer-readable medium, which is further taught per Griffin’s FIG. 1 and [0023]-[0029] (e.g., the taught “health evaluator” includes memory, control circuitry, etc. that read on these further recitations).
11. Claim 4 is rejected under 35 U.S.C. 103 as being unpatentable over Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 and further yet in view of Non-Patent Literature “Patient Outcomes After Applied Behavior Analysis for Autism Spectrum Disorder” (“Choi”).
Regarding claim 4, Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 teach the method of claim 1, as discussed above. The instant claim further recites the additional limitations wherein the data associated with the subject comprise the schooling data, wherein the schooling data comprise an indication of whether the subject attends school, an indication of whether the subject has been assigned a school aide, an indication of whether the subject is a part of a special education program, or combinations thereof. Regarding this, Griffin in view of Kohli teaches the consideration of a patient’s “educational background” (Kohli’s page 20, 2nd column, 1st full paragraph: “Further, sociodemographic characteristics such as age, gender, place of residence, access to healthcare, family income, and educational background can affect the treatment design and delivery. The above challenges can be overcome by designing a feature vector during the patient intake to capture diagnostic and functional assessment scores, age, gender, and other sociodemographic characteristics. At the intake stage, using the feature vector, the patient similarity model can compare incoming patients to an extensive patient database to recommend the most similar patients and correlate their treatment trajectory with outcomes, allowing physicians to select the ideal treatment strategy”), which could indicate “whether the subject attends school” (as recited in the instant limitation).
However, to the extent that Kohli’s teaching referenced above is not deemed sufficient, the Examiner further relies upon a more firm teaching found in Choi, see e.g., page 5, under the heading Predictor Variables and Covariates: “Our predictor variables were ABA dose and service history (past and current receipt of ABA, past and current receipt of special education ...” (which concretely reads on the recitation for “an indication of whether the subject is a part of a special education program”).
Like Kohli, Choi is directed to an evaluation of patient information in EHR/EMR data to identify patterns that are predictors for meaningful predictions for a similar subject population. Hence, they are similarly directed and therefore analogous. It would have been obvious to incorporate consideration of therapeutic treatment, as Choi specifically contemplates, in the same/like manner that Kohli and Griffin already more broadly/generally do, with a reasonable expectation of success, as a way to incorporate an additional piece of information that could serve as a meaningful predictor (as Choi purports) to further improve the model performance per Griffin and Kohli.
12. Claim 11 is rejected under 35 U.S.C. 103 as being unpatentable over Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 and further yet in view of Non-Patent Literature “How to Tune the Number and Size of Decision Trees with XGBoost in Python” (“Brownlee 2”).
Regarding claim 11, Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 teach the method of claim 10, as discussed above. The instant claim further recites the additional limitation wherein the gradient-boosted tree model comprises from 50 decision trees to 400 decision trees, and wherein the NDDTR model has a tree depth of at least 2 and not more than 6. The aforementioned references teach the use of gradient-boosted tree models (e.g., as discussed per claim 10, Griffin’s [0050] for example) and make some clarification that such tree models would have a “number of layers” per Griffin’s [0053] (which the Examiner equates with “tree depth”). That said, neither Griffin nor Kohli present a clear teaching as to the number of trees and a numerical representation or range thereof for tree depth. Rather, the Examiner relies upon BROWNLEE 2 to teach what Griffin etc. otherwise lack, see e.g., Brownlee 2’s page 3 (contemplates a tunable best number of trees in a range spanning from 100 to 350) and page 4 (contemplates a tunable depth between 1 and 9). Respectfully, the ranges for tree number and tree depth as taught by the reference overlap with the ranges for both aspects as recited, therefore providing a teaching that can read on the limitation as recited.
Like Griffin, Brownlee 2 is directed to a tree-based modeling approach. Hence, they are similarly directed and therefore analogous. It would have been obvious to incorporate Brownlee 2’s tunable aspects for its tree structures and model with Griffin’s modified framework, with a reasonable expectation of success, such as a way to permit tuning of the model as one generally does with model tuning to promote model efficiency, accuracy, and so forth.
13. Claim 12 is rejected under 35 U.S.C. 103 as being unpatentable over Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 and further yet in view of Non-Patent Literature “Goal-Oriented Sensitivity Analysis of Hyperparameters in Deep Learning” (“Novello”).
Regarding claim 12, Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 teach the method of claim 10, as discussed above. The aforementioned references teach the further limitation for identifying NDDTR model hyperparameters, wherein the NDDTR model hyperparameters comprise tree depth, number of decision trees, learning rate, scale positive weight, alpha regularization parameter, gamma regularization parameter, or combinations thereof (Griffin’s discussion of model hyperparameters, inclusive of “tree layers” (i.e., “tree depth” as recited), per [0053]) and certainly using the taught hyperparameters for tuning the NDDTR model hyperparameters ... (Examiner: that is the point of hyperparameters), but not the further limitation wherein the tuning of the NDDTR model hyperparameters is effective to provide for an NDDTR model sensitivity of from 0.75 to 0.99. Rather, the Examiner relies upon NOVELLO to teach what Griffin etc. otherwise lack, see e.g., Novello’s page 20, section 5.1.1, discussing an accuracy going up to about 99%.
Like Griffin, Novello is directed to a comparable modelling approach that involves hyperparameter tuning of the model. Hence, they are similarly directed and therefore analogous. It would have been obvious to incorporate Novello’s sensitivity analysis with Griffin’s modified framework, with a reasonable expectation of success, to promote accuracy in the model as Novello teaches with its sensitivity analysis.
14. Claims 17 and 19-22 are rejected under 35 U.S.C. 103 as being unpatentable over Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee and further yet in view of WO 2022051674 A1 (“Becich”).
Regarding claim 17, the claim includes the same or similar limitations as claim 1 discussed above, and is therefore rejected under the same rationale. The instant claim specifically recites many of the same model aspects but with respect to its training and essentially its creation, which the Examiner believes Griffin sufficiently reads on, see e.g., Griffin’s Abstract (“A predictive patient health machine learning model is trained based on baseline health data configured as directed graphs. Patient-healthcare system encounter data formed at least in part by electronic medical records (EMRs) is gathered. The patient-healthcare system encounter data is configured as directed graphs to generate graphed health data and the predictive patient health machine learning model is trained on that graphed health data.”), with [0002], [0008], [0017], and [0021]-[0023] for example serving as restatements of the Abstract’s substance.
The present claim additionally recites a limitation for reducing the dimensionality of training data … which the aforementioned references Griffin etc. as discussed per claim 1 for example do not teach. Rather, the Examiner relies upon BECICH to teach what Griffin etc. otherwise lack, see e.g., Becich’s [00119] generally discussing iterative training and more specifically [00313] discussing discussing sequential removal of features during feature selection as part of cross validation, based on optimization using AUROC, which the Examiner equates with forward selection dimensionality reduction (i.e., recited (iii) of the possible ways to reduce dimensionality of training data as recited in the present claim).
The claim additionally conditions the model’s implemented operating threshold to prioritize true positives and limit false negatives, e.g. as the Examiner specifically addresses per claim 1, as while testing the NDDTR model on a hold-out test set, which the Examiner believes Becich teaches: its page 8, 1st column, bottom paragraph discussing “We used 80% of user and item vectors for the training algorithm and masked 20% of items to blind validate model performance.”, where the Examiner reasons that the 20% set aside for validation as taught is equivalent to the recited hold-out test set.
Becich provides a validation approach to improving its modelling, which correlates with Kohli’s approach (section 3.1’s trial objectives, noted as (3): “to develop and validate machine learning models to recommend and personalize behavioral treatment plans using longitudinal treatment data of children across multiple domains, such as expressive language, receptive language, echoic, and requesting.”). Hence, Becich is directed to the same modelling considerations as Kohli and is applicable to a modification of Griffin in view of Kohli. It would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate Becich’s dimensionality reduction into the modified framework of Griffin etc. (per the Examiner’s discussion of claim 1) as a means of providing cross validation advantages that improve the model’s performance as it is trained and optimized.
Regarding claim 19, Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee and further yet in view of Becich teach the method of claim 17, as discussed above. The aforementioned references teach the further limitations wherein the feature selection method further comprises (1) evaluating the area under the receiver operator characteristic curve (AUROC) of each single data feature (Becich’s [0005] discussing AUROC in relation to improved performance and sensitivity); (2) removing data features that yield single feature AUROC values of equal to or less than 0.55 (Becich’s AUROC threshold appears to be at least 0.70 if not higher (based on [0027]-[0031] and [0034]-[0038]), hence it stands to reason that values below that would not be subject to selection); (3) evaluating the AUROC of the combined remaining data features and (4) iteratively training the NDDTR model by removing one data feature at a time with replacement from the data features remaining in the training dataset (Becich’s [00119] generally discussing iterative training and more specifically [00313] discussing forward selection iteration), wherein the NDDTR model is trained using cross-validation, and wherein feature subsets are not reshuffled between folds (Becich’s [00345] discussing an approach using folds that is silent to reshuffling for purposes of cross validation). The motivation for combining the references is as discussed above in relation to claim 17.
Regarding claim 20, Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee and further yet in view of Becich teach the method of claim 19, as discussed above. The aforementioned references teach the additional limitations further comprising eliminating one or more of the data features causing the highest increase in mean cross-validation AUROC when removed (Becich’s [00313] discussing sequential removal of features during feature selection as part of cross validation, based on optimization using AUROC). The motivation for combining the references is as discussed above in relation to claim 17.
Regarding claim 21, the claim includes limitations similar to that as discussed above in relation to claim 1, and is therefore rejected under the same rationale. Specifically, see the Examiner’s citation to Griffin’s [0050], discussing that the health predictor model can be an ensemble model of decision trees, as part of a larger discussion of the tree based model per [0050]-[0060] (thereby reading on the limitation that the model may consist of “a decision tree type model” as recited). Hence, the present claim is appropriately rejected under Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee and further yet in view of Becich, with a motivation to combine the references as discussed above in relation to claim 17.
Regarding claim 22, Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee and further yet in view of Becich teach the method of claim 21, as discussed above. The instant claim further recites the additional limitations wherein the machine learning model is a gradient-boosted tree model comprising a plurality of weighted decision trees (Griffin’s [0050]: “... The health predictor model can generate a final prediction based on individual predictions made by multiple classification models forming the health predictor model. In a specific example, the health predictor model is formed based on gradient boosted decision trees. The health predictor model can be based on parallel decision tree boosting. In one example, the health predictor model utilizes the XGBoost algorithm during training of the health predictor model.”). The motivation for combining the references is as discussed above in relation to claim 17.
15. Claim 13 is rejected under 35 U.S.C. 103 as being unpatentable over Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 and also Novello and further yet in view of Brownlee 2 and CN 113609841 (“Li”).
Regarding claim 13, Griffin in view of Kohli and further in view of Datasklr, Scikit, and Brownlee 1 and also Novello teach the method of claim 12, as discussed above. The aforementioned references teach the entirety of the further limitations, and hence the Examiner relies upon further references Brownlee 2 and LI to teach what they lack:
A further limitation wherein the gradient-boosted tree model comprises from 50 decision trees to 150 decision trees (as discussed per claim 10, Griffin’s [0050] for example teaches the use of gradient-boosted trees, but is silent as to how many, and rather the Examiner relies upon Brownlee 2 to teach that, see e.g., Brownlee 2’s page 3 (contemplates a tunable best number of trees in a range spanning from 100 to 350)).
A motivation for modifying Griffin in view of Brownlee 2 for this same/similar limitation has been discussed above in relation to claim 11.
A further limitation wherein the NDDTR model has a tree depth of not more than 3 (Griffin’s tree models would have a “number of layers”, per Griffin’s [0053] (which the Examiner equates with “tree depth”), but is silent as to how many layers, and rather the Examiner relies upon Brownlee to teach that, see e.g., Brownlee 2’s page 4 (contemplates a tunable depth between 1 and 9));
A motivation for modifying Griffin in view of Brownlee 2 for this same/similar limitation has been discussed above in relation to claim 11.
A further limitation wherein the NDDTR model has a learning rate of equal to or less than 0.4 (learning rate/parameter taught by Li, page 11: “parameter learning-rate=0.01, the proper adjustment of the parameter is good for improving the precision of the model”);
Like Griffin etc., Li is directed to creating a model and tuning it appropriately via various parameters, thereby promoting accuracy and performance in use of the model. Hence, they are similarly directed and therefore analogous. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to incorporate Li’s tuning aspects, as discussed here, into Griffin’s modified framework, with a reasonable expectation of success, so as to improve model performance and accuracy.
A further limitation wherein the NDDTR model has a scale positive weight of from 0.1 to 10 (weighting according to scale as taught by Li, page 11: “parameter scale-pos-weight t= 1, setting this value is because the class is not balanced”);
A motivation for modifying Griffin in view of Li for this same/similar limitation has been discussed just above.
A further limitation wherein the NDDTR model has an alpha regularization parameter of from 0 to 1 (gradient of alpha of the learning parameter as taught by Li, page 11: “parameter reg-alpha=0.005, the parameter represents the weight of L1 regularization item, applying in the condition of higher dimension, the speed of the model is faster”); and
A motivation for modifying Griffin in view of Li for this same/similar limitation has been discussed just above.
A further limitation wherein the NDDTR model has a gamma regularization parameter of from 0 to 1 (a gramma regularization taught by Li, page 11: “parameter gamma=0, when the node is split, only the value of the loss function after splitting is reduced, then splitting the node. The gamma specifies a minimum loss function drop value required for node splitting. The larger the value of the parameter, the more conservative model. The value of this parameter is associated with the loss function.”).
A motivation for modifying Griffin in view of Li for this same/similar limitation has been discussed just above.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SHOURJO DASGUPTA whose telephone number is (571)272-7207. The examiner can normally be reached M-F 8am-5pm CST.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Tamara Kyle can be reached at 571 272 4241. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/SHOURJO DASGUPTA/Primary Examiner, Art Unit 2144
Prosecution Insights